In this week’s Friday Faves, we have Tensorflow 2 dropping and a beautiful bonus, next level physics-based ML, and a problem with a Harvard deep learning paper.
Tensorflow 2 for Researchers
Tensorflow 2.0 dropped this week and it has Eager Execution (read “normal behaviour”) per default and the Keras API per default.
If you’re familiar with PyTorch, you’ll wonder “that’s news?” but you can now use one of three, which for PyTorch users will be another “that’s news?”
Depending on how reproducible and complicated your model gets, you’ll need to make that choice. I have personally yet to use Subclassing, but I hear good things. Specifically, good things by François Chollet, the creator of Keras and Google employee. He compiled a truly amazing thread for researchers:
If you’re too busy, you can go straight to the Google Colab. But I really recommend looking at the thread, there are some gems in there.
If you’ve been around any geophysics or machine learning conference, you have seen the hype of physics-based machine learning. And personally, I really like the interesting combinations of known methods with neural networks. So DeepMind swoops in and combines some interesting concepts, namely Hamiltonian physics, graph neural networks and a sprinkling of Runge Kutta.
The Aftershock to the Aftershock AI
Earlier 2019 Google and Harvard published a paper in Nature, predicting aftershock patterns with a deep neural network. It did not take long for people to poke holes in the paper. It did not take long for the Harvard scientists to respond indignantly. It did, however, take a bit to get the following preprint published in Nature Matters Arising:
One neuron versus deep learning in aftershock prediction
https://arxiv.org/abs/1904.01983 and the Nature paper is available here: https://www.nature.com/articles/s41586-019-1582-8.epdf
It shows the problems of Deep Learning being terribly publishable, but basic data science principle not being followed, aka: did you try a simpler model?